Howie is the Claude Opus 4.7 character who watches every active shift, drafts incident narratives, and ranks responders — so the human dispatcher gets a one-paragraph brief instead of a dashboard to interpret. Howie recommends. Humans decide.
Howie is built on Claude Opus 4.7 — chosen because the long-context reasoning means the model can hold an entire 12-hour shift in working memory and notice the deviation a dashboard would never surface.
Howie scans every active worker — last GPS ping, last welfare ack, vitals from connected wearables, expected check-in cadence — and flags anyone drifting from baseline. Dispatchers see a single ranked list, not 200 dots on a map.
The moment a worker triggers SOS, Howie writes a one-paragraph situation brief: who, where (with what3words + nearest postcode), what the wearable saw in the 60 seconds before the trigger, who the nearest responders are. Dispatchers read it before the call connects.
Claude re-ranks Howie's GPS shortlist with rationale — "Sam is 230m closer but off-shift, Priya is on-shift and first-aid trained". The dispatcher picks; Howie records the why, so the audit log explains every decision.
An incident can't be closed until three things happen: responder confirms on-scene, worker confirms safe (or alternative authority confirms welfare), and a named dispatcher signs off. Howie blocks premature closes — even from admins.
Open /ai-monitor in admin and you see Howie's live thought stream — every welfare check, every shortlist, every flagged anomaly — alongside the human dispatcher action that confirmed or overrode it. Every row links back to the incident, the worker, and the underlying wearable telemetry.
We've seen too many AI safety stories that started with “the model decided to…” Howie can't. The runtime physically cannot perform any of these three actions without a named human acknowledgement.
All outbound calls are placed by a named human dispatcher (or the worker's own device). Howie can recommend the call and pre-fill the script — but the human hits dial.
Only a named human with the right role can close. Howie can mark an alert as "likely resolved — Mei acked safe at 14:21" but the close action stays human-only.
If a dispatcher disagrees with Howie's responder ranking, the human pick is the answer. Howie records the disagreement for later review — not for punishment, for learning.
Each model output carries a deterministic ID, the prompt hash and the dispatcher who acked it. If you're asked “why did this happen?” six months later, the answer is one query away.
No per-seat AI surcharge, no per-token billing. Howie's Claude usage is metered against a fair-use envelope sized to your active-worker count — and visible from the Trust + Verify panel in admin so you can see exactly what's being spent.
Spin up a demo tenant and trigger a synthetic SOS — you'll see Howie's brief in your inbox within seconds.